Micron CEO Says AI Still in 'Early Stages' as Memory Supply Tightens

Micron Technology posted record-breaking results in its fiscal second quarter, as CEO Sanjay Mehrotra declared that the current AI wave remains in its 'early stages' — and that memory supply constraints show no signs of easing. The storage giant now forecasts that AI-driven demand for DRAM and NAND flash will exceed 50% of the total addressable memory market in 2025, a milestone that underscores just how profoundly artificial intelligence is reshaping the semiconductor industry.

Mehrotra's comments come at a pivotal moment for the memory sector. While AI infrastructure buildouts accelerate across hyperscalers and enterprise data centers, the supply of high-performance memory chips remains critically tight — creating both opportunity and urgency for Micron and its competitors Samsung and SK Hynix.

Key Takeaways

• Micron posted multiple all-time records in fiscal Q2 and expects to break them again in Q3
• CEO Mehrotra says AI is still in its 'early stages,' with inference workloads reaching an inflection point
• AI demand for DRAM and NAND is projected to surpass 50% of total industry market size this year
• HBM3 yields are maturing, while HBM4E memory is on track for mass production next year
• Consumer electronics remain weak — PC and mobile device sales expected to decline by low double digits
• Memory is now considered a 'strategic asset,' prompting Micron to invest heavily in global manufacturing capacity

AI Inference Hits an Inflection Point, Driving Unprecedented Memory Demand

Mehrotra emphasized that the AI industry is entering a new phase where inference workloads — the process of running trained models to generate outputs — are becoming the dominant driver of memory consumption. Unlike training, which happens in centralized clusters, inference is distributed across a widening range of deployments, from cloud APIs to edge devices.

The CEO specifically pointed to token generation as a key factor. As large language models produce longer, more complex outputs and handle multi-step reasoning chains, the demands on memory speed and capacity grow exponentially. Each token generated requires rapid access to model weights stored in memory, and as models scale to hundreds of billions of parameters, the bandwidth and capacity requirements become staggering.

This inference inflection point means that memory is no longer just a commodity component — it is a performance bottleneck. The speed at which a GPU can generate tokens is directly constrained by how fast it can read from and write to memory. Mehrotra described memory as a 'strategic asset,' a term that signals how the industry's perception of DRAM and NAND has fundamentally shifted in the AI era.

HBM Roadmap Accelerates to Keep Pace with GPU Evolution

On the technology front, Micron is aggressively advancing its High Bandwidth Memory (HBM) portfolio to meet the demands of next-generation AI accelerators. HBM chips are stacked vertically and mounted directly alongside GPUs, providing the massive bandwidth needed for AI training and inference.

Micron reported that its HBM3 products have reached mature manufacturing yields, meaning the company can now produce them reliably at scale. This is critical because HBM manufacturing is notoriously difficult — the process of stacking multiple DRAM dies using through-silicon vias (TSVs) has historically suffered from low yield rates that constrain supply and inflate costs.

Looking further ahead, Micron confirmed that HBM4E memory is on track for volume production next year. HBM4E represents the next major leap in bandwidth and capacity, designed to pair with upcoming GPU architectures from NVIDIA and AMD. The 'E' designation typically indicates an enhanced version with higher density or improved performance characteristics.

• HBM3: Yields now mature, enabling scaled production for current-generation AI accelerators
• HBM3E: Currently shipping to major customers including NVIDIA for its H200 and B200 GPUs
• HBM4: Next-generation standard with a new base-die architecture
• HBM4E: Mass production slated for 2026, targeting next-gen GPU platforms

The HBM market has become fiercely competitive. SK Hynix currently leads in HBM3E market share, while Samsung has been working to close the gap after earlier yield challenges. Micron's ability to mature HBM3 yields and stay on track with HBM4E positions it as a credible third player in what has become one of the most strategically important segments of the semiconductor industry.

Supply Constraints Persist as Capacity Expansion Proves Difficult

Perhaps the most consequential aspect of Mehrotra's commentary was his warning about supply tightness. Despite strong demand signals, the memory industry cannot simply flip a switch to increase output. Building new fabrication facilities takes 2 to 3 years, and the advanced packaging processes required for HBM add additional complexity.

Micron indicated that industry-wide supply of both DRAM and NAND remains constrained, and that expanding capacity is neither quick nor cheap. A single advanced memory fab can cost upward of $10 billion to build, and the specialized equipment needed — particularly extreme ultraviolet (EUV) lithography tools from ASML — faces its own supply limitations.

This supply-demand imbalance has significant pricing implications. Memory prices, which crashed during the 2022-2023 downturn, have recovered sharply as AI demand absorbs available supply. Analysts expect this favorable pricing environment to persist through at least the end of 2025, benefiting all 3 major memory makers.

Micron is responding by investing in global manufacturing expansion. The company has committed billions to new facilities in the United States, including a major fab project in Clay, New York, supported by CHIPS Act funding. It is also expanding operations in Japan and India to diversify its manufacturing footprint.

LPDDR and DDR5 Power the Server and Mobile AI Stack

Beyond HBM, Micron is capitalizing on strong demand for LPDDR (Low Power DDR) memory in both server and mobile applications. LPDDR has emerged as the preferred memory type for AI inference servers because it offers a favorable balance of bandwidth, capacity, and power efficiency compared to standard DDR.

Micron showcased high-capacity LPCAMM2 modules — a newer form factor that enables easier upgrades and higher density in laptops and workstations. The company is also supplying DDR5 memory to NVIDIA for its server platforms, ensuring compatibility with the latest data center GPU architectures.

In the server market specifically, the shift to DDR5 is accelerating. Every new AI server platform from NVIDIA, AMD, and Intel now requires DDR5, and the higher per-server memory capacity needed for AI workloads means each unit consumes significantly more DRAM than traditional servers.

• LPDDR5X: Preferred for AI inference servers due to low power consumption
• LPCAMM2: New modular format enabling higher-density laptop memory configurations
• DDR5: Now standard for all next-generation server platforms
• CXL-attached memory: Emerging technology that could further expand addressable memory pools

Consumer Electronics Weakness Contrasts with AI Strength

Not all segments are benefiting equally from the AI boom. Micron expects PC and smartphone sales to decline by low double-digit percentages this year, reflecting broader consumer spending caution and longer device replacement cycles. This weakness in consumer electronics partially offsets the strength in AI-related demand.

However, Mehrotra noted an important trend that could eventually lift the consumer segment: the rise of on-device AI. As AI agent workflows begin running locally on PCs — handling tasks like document summarization, code generation, and workflow automation — the memory requirements for mainstream computers are increasing dramatically.

Micron believes that 32GB of RAM will become the standard configuration for AI-capable PCs, up from the 8GB or 16GB that has been typical for years. This doubling or quadrupling of per-device memory content represents a significant growth opportunity, even if unit sales remain flat or decline. If the average PC ships with twice as much DRAM, total memory demand from the PC segment could grow even as fewer units sell.

What This Means for the AI Industry

Micron's results and outlook carry broad implications for the AI ecosystem. For cloud providers and AI startups, the message is clear: memory will remain expensive and potentially supply-constrained for the foreseeable future. Companies planning large-scale AI deployments should factor in memory availability as a potential bottleneck alongside GPU supply.

For investors and analysts, Micron's characterization of AI as 'early stage' suggests that the memory demand cycle has significant room to run. Unlike previous semiconductor cycles that peaked and crashed quickly, the structural nature of AI demand — driven by inference scaling, model growth, and new use cases — could support a longer upcycle.

For hardware engineers and system architects, the rapid evolution of memory standards means that design decisions made today will need to account for HBM4, LPDDR6, and CXL memory technologies that are already in development. The gap between memory bandwidth and compute capability remains one of the most critical challenges in AI system design.

Looking Ahead: Micron Expects Another Record Quarter

Mehrotra expressed confidence that Micron will post yet another record-breaking quarter in fiscal Q3, driven by continued strength in AI-related memory demand. The company's ability to execute on its HBM roadmap while maintaining disciplined supply will be key to sustaining this momentum.

The broader narrative is one of structural transformation. Memory, once viewed as a cyclical commodity prone to boom-and-bust cycles, is being repositioned as a strategic technology asset in the AI era. As long as the appetite for AI training and inference continues to grow — and Mehrotra's 'early stages' comment suggests it will — companies like Micron stand to benefit from a demand environment unlike anything the industry has seen before.

The key risk remains macroeconomic uncertainty. Trade tensions, tariffs, and potential consumer spending slowdowns could dampen the non-AI portions of Micron's business. But for now, the AI tailwind appears strong enough to carry the company — and the entire memory industry — to new heights.